Remove Image Backgrounds Locally with WebAssembly

Removing a background is not edge detection and it is not colour keying. The task a modern remover performs is salient object detection: predicting, for every pixel, how likely it is to belong to the visually dominant subject.

The model used here is U²-Net in its small variant, u2netp. U²-Net is a nested U-structure — an encoder–decoder built from further encoder–decoder blocks — designed specifically for salience rather than for classification. It does not know what a dog is. It has learned which regions of an image human annotators marked as the subject.

The output is not a mask of ones and zeros. It is a saliency map: a single-channel image the same shape as the input, where each value is a confidence between 0 and 1. That map becomes the alpha channel of the result, which is why good removers produce soft, believable edges on hair and fur rather than the jagged cut-out that thresholding produces.

The small variant matters for a browser. u2netp is a few megabytes rather than the ~170 MB of the full network, which is the difference between a tool that loads in seconds and one nobody waits for.

U²-Net takes a fixed input of 320×320 pixels. Every image, whatever its dimensions, is resized to that square before inference and the resulting saliency map is scaled back up to the original size.

This single fact explains most of the quality characteristics people notice:

• Fine detail below the resample threshold is lost. A 4000-pixel-wide photograph is reduced to 320 before the model sees it, so individual hair strands, wire-frame glasses and chain-link fences fall below one pixel of model resolution. The upscaled matte approximates them rather than resolving them.
• Extreme aspect ratios suffer. A panorama squeezed into a square distorts the subject before inference, and accuracy drops.
• Cost is constant. Inference on a 500-pixel image and a 5000-pixel image takes the same time, because both become 320×320. Only the decode and the final compositing scale with image size.

Before inference the pixels are also normalised: converted to floating point, divided by 255, then standardised against the mean and standard deviation the network was trained on. Skipping that step produces a saliency map that looks like noise — a common failure when people wire up the model themselves and feed it raw bytes.

A tool that never says what it cannot do wastes people's time. Salient object detection has predictable failure modes:

Low subject–background contrast. A grey cat on a grey sofa gives the model little to separate. The matte becomes uncertain in exactly the region that matters.

Multiple plausible subjects. The model predicts *the* salient object. Three people standing apart yields an arbitrary choice or a blended matte covering all of them.

Transparency and translucency. Glass, smoke, water, veils and fine netting have no correct binary answer. A wine glass returns either a solid silhouette or a hole, and neither is right.

Very fine structure. Flyaway hair, fur at the edge of a backlit subject, thin cables and foliage. The 320-pixel constraint sets a hard floor here.

Subjects that touch the frame edge. Salience models are trained predominantly on centred subjects; something cropped at the boundary is often partially excluded.

For product photography on a plain backdrop the results are usually publication-ready. For a backlit portrait with loose hair, expect to do manual work afterwards — and choose a tool that gives you the matte rather than only a flattened result.

Background removal is applied overwhelmingly to photographs of people, identity documents, and unreleased product shots. The standard workflow for every major online remover is to upload the original image to a server, run inference there, and return the cut-out.

That means the original — full resolution, with its EXIF metadata intact, including GPS coordinates and camera serial number where present — has been transmitted to and processed by a third party. For a passport photograph or an unannounced product, that is a meaningful disclosure, and it is invisible to the user because the interface looks identical either way.

Running inference in the browser removes the transfer entirely. The model file is fetched once, cached, and executed by ONNX Runtime Web, which uses WebAssembly and, where available, WebGPU. The image is decoded into a canvas, converted to a tensor, and passed to the model — all inside the tab.

The honest trade-offs of the local approach:

• The first run downloads the model. Subsequent runs use the cached copy.
• It uses your device's compute. On an older phone a large image takes noticeably longer than a server would.
• Quality is bounded by the small model. Services running the full U²-Net or a proprietary successor on server GPUs can produce better mattes on difficult images.

The OpenTools background remover makes the local trade deliberately: the image never leaves the device, and the worker running the model is terminated when you navigate away so the multi-megabyte model is not left resident.

The model is fixed, but the input is not, and input quality dominates the result.

Shoot or crop for contrast. The single highest-leverage change is separating the subject from the background tonally. A dark subject against a light backdrop is trivial; a dark subject against a dark backdrop is unreliable.

Fill the frame, but leave margin. Subjects touching the frame edge are often clipped by the salience prediction. A little breathing room around the subject measurably improves the matte.

Do not pre-sharpen. Sharpening amplifies edge halos, which the model can read as subject boundary, producing a fringe of background pixels retained around the cut-out.

Composite onto a similar tone. If the final background is dark, any retained fringe from a light original will glow. Where you control the destination, matching the tone hides small matte errors that would otherwise be obvious.

Export PNG or WebP, never JPEG. JPEG has no alpha channel. Exporting a cut-out as JPEG flattens the transparency onto white and discards it permanently. This is the most common way people lose a good result after producing it.